We used a dataset already used in the ACM DEBS 2013 Grand Challenge \cite{Mutschler_2013}, a competition of soccer analysis. This dataset has been collected by the Real-Time Locating System deployed on a football field of the Nuremberg Stadium in Germany. Data originates from sensors located near the players’ shoes (1 sensor per leg) and in the ball (1 sensor). The goal keeper is equipped with two additional sensors, one at each hand. The sensors in the players’ shoes and hands produce data with 200Hz frequency, while the sensor in the ball produces data with 2000Hz frequency. The total data rate reaches roughly 15.000 position events per second. Every position event describes position of a given sensor in a three-dimensional coordinate system. Furthermore this dataset provides both positions and accelerations, which are the data we were looking for.
For our experiments we used only data during a restricted period of several minutes beginning at arbitrary time of the play. Computationnaly, a trajectory will be defined as a set of 2-dimensional positions and accelerations regularly distributed in time. Furthermore we will handle a set of synchronized trajectories. This synchronization is highly important since we will assume it when we will approximate the joint probability distribution of positions or accelerations of two players.
Data are read and extracted with a 25Hz frequency which corresponds to the best frequency we could have if we were extracting the trajectories directly from the videos of the soccer match. As each player is equipped with several sensors we combine their data to have one and only one trajectory per player. Sometimes some data have been missed, but one can cope with this problem interpolating data so as to maintain synchronization between the different trajectories.
When a player goes out of the field we can’t neither treat it as usual with positions out of the field because it would be non sense nor stop extracting the data because it would break the continuity of extracted player’s position. Whereas the displacement of a player going and fetching the ball out of the field is non sense, the paths of other players at this while bring relevant information. Thus in this case we just consider that the outside player remain at the last read position on the field with a null acceleration.
We also decide to exclude the ball trajectory because it very often goes out of the play and also because it occurs a lot of ball change.